Seminar given by Bill Bialek of Princeton University to the Redwood Center for Theoretical Neuroscience at UC Berkeley on July 23, 2008.
In this informal talk I'll try to give a feeling for three problems my colleagues and I are thinking about. A bit of a hodge-podge, perhaps, but I hope that the combination of topics provokes discussion: (1) How do we go from what we can measure about neurons to a global picture of the network dynamics? We've been exploring maximum entropy methods that allow us to construct a statistical mechanics of the network from measurements on correlations between pairs of neurons, as well as more direct paths to construct a thermodynamics of the network. The surprise emerging from this analysis is that a real neural network (the retina responding to naturalistic movies) seems to be poised at a critical point. (2) How can the brain calibrate the neural code without access to independent knowledge of the stimulus? Rather than asking how patterns of neural activity are related to sensory stimuli in the recent past, we have been exploring how this activity is related to activity in the immediate future. This problem of extracting predictive information is quite general (and quite rich), and seems to identify patterns which are especially informative about the sensory input even though the analysis makes no reference to these inputs; again our example is the retina, where we can also try to understand the nature of the stimulus features that are detected by the prediction algorithms. (3) Can we understand the architecture (or even the detailed dynamics) of networks as solutions to some optimization problem for the flow or representation of information? In the context of neurons, this is an old idea; it has an interesting mapping to genetic networks, where the physical constraints on information flow are clearer. We have some surprising initial successes in applying such optimization principles to the first steps of genetic control in the development of the fruit fly embryo, and I'll try to highlight analogies between current questions in neural and genetic circuits.
Note about the video recording: About 1 minute of sound is missing at the start of the question answer session (about 1 hour 10 minutes into the talk) and then the sound volume is reduced (but present).